Machine learning based prediction of waste activated sludge generation for optimization of WWTP operational efficiency

This study predicted the amount of waste activated sludge withdrawn for external disposal based on process operation and water quality data using machine learning models. By combining hyperparameter optimization with a sliding-window–based moving average method, the performances of Random Forest, XGBoost, and LightGBM models were compared and analyzed. Additionally, we present an integrated prediction–optimization pipeline that couples the model-predicted waste activated sludge with NSGA-II to provide the optimal SRT under effluent water-quality constraints. The analysis showed that the XGBoost-Exp1 model exhibited the best predictive performance with an R² of 0.911, an RMSE of 87.59, and an MAE of 65.25. SHAP analysis revealed that COD_eff and COD_inf variables were closely related to actual waste activated sludge generation in field operations and made the greatest contribution to prediction. NSGA-II produced a Pareto frontier of SRT set-points meeting effluent constraints, enabling regulation-compliant operating choices that balance WAS reduction and effluent performance. This study demonstrated that the proposed modeling strategy maintains high predictive accuracy under various operating conditions, confirming its potential to enhance WWTP operational efficiency and inform predictive maintenance strategies. Moreover, the proposed model suggests feasibility for integration into real-time predictive systems for process automation.